Docker – running mesos-master in docker container – cannot see containers IP from other hosts

You don't know the IP address before Marathon finds suitable host for the task. However, at the time of launching a task at Mesos slave some properties are stored in ENV variables:

...
ULIMIT=-n 8192
HOST=slave01.example.com
MESOS_SLAVE_PID=slave(1)@192.168.1.2:5051
MARATHON_APP_RESOURCE_CPUS=0.1
MESOS_TASK_ID=dummy.eb6e1d9a-c2f0-11e5-a58b-00259057db2f

If you're fine with hostname, you can use HOST variable or extract the IP address from MESOS_SLAVE_PID.

If you start mesos-slave with --ip=192.168.1.2, then both HOST and LIBPROCESS_IP ENV variable will be set to host's address. For Mesosphere package:

echo "192.168.1.2" > /etc/mesos-slave/ip

and restart slave.

Yet another option is to resolve agent's IP inside Docker container:

AGENT_IP=$(getent hosts ${HOST} | awk '{ print $1 }')

It takes some time to fetch a Docker image, start container and finally start serving requests. Ideally you should reload HAProxy by the time your new instance is ready. But there might be still active connections to your old instance. When you start investigating this issue it turns out that a solution already exists, it's called blue-green deployment as described by M. Fowler.

If you aim for zero-downtime, there won't be a trivial solution. Yelp engineers describe how to reload HAProxy with true zero downtime by adding extra rules to iptables.

Recently blue-green deployment scheme was merged into marathon-lb (including the iptables trick from Yelp). I think that consul-mesos currently doesn't support this.